一、处理原始数据集¶
将 kaggle 提供的训练集、测试集压缩文件下载到本地后，进行解压缩。将所有的训练集图片放在 train 文件夹中，测试集的图片都放在 test 文件夹中。
In [1]:
import os
import zipfile
from os.path import isfile, isdir
if isfile("train.zip"):
if not isdir("train"):
with zipfile.ZipFile("train.zip","r") as zf:
zf.extractall()
zf.close()
if isfile("test.zip"):
if not isdir("test"):
with zipfile.ZipFile("test.zip","r") as zf:
zf.extractall()
zf.close()
if isdir("train"):
print("train images num: %d" % len(os.listdir("train")))
if isdir("test"):
print("test images num: %d" % len(os.listdir("test")))
In [2]:
#随机预览几张图片
from keras.preprocessing.image import *
import matplotlib.pyplot as plt
%matplotlib inline
%config InlineBackend.figure_format = 'retina'
fig = plt.figure(figsize=(20,8))
for i in range(14):
a = fig.add_subplot(2, 7, i+1)
if i < 7:
file = "cat.%s.jpg" % np.random.randint(12500)
else:
file = "dog.%s.jpg" % np.random.randint(12500)
a.set_title(file)
a.axis('off')
img = load_img("train/"+file)
a.imshow(img, interpolation="nearest")
In [3]:
import numpy as np
from PIL import Image
from tqdm import tqdm
shape_list = []
for file in tqdm(os.listdir('train')):
img = Image.open("train/"+file)
x = img.size
shape_list.append((x[0], x[1]))
In [4]:
from collections import Counter
import matplotlib.pyplot as plt
shape_collect = Counter(shape_list)
x_w = []
y_h = []
area = []
for k in shape_collect:
area.append(shape_collect[k])
x_w.append(k[0])
y_h.append(k[1])
colors = np.random.rand(len(x_w))
plt.scatter(x_w, y_h, s=area, c='r', alpha=0.5)
plt.title('image size distribution')
plt.show()
In [5]:
tiny_pics, large_pics = [], []
for file in tqdm(os.listdir('train')):
img = Image.open("train/"+file)
x = img.size
if x[0] < 100 and x[1] < 100:
tiny_pics.append(file)
elif x[0] > 800:
large_pics.append(file)
print("tiny pics: %s\n%s" % (len(tiny_pics), tiny_pics))
print("large_pics: %s\n%s" % (len(large_pics), large_pics))
二、整理文件结构¶
利用 keras 预训练模型筛选出训练集中的异常值。keras 的 ImageDataGenerator 需要重新安排文件结构，最终需要将文件夹结构设置为：
data/
train/
dogs/
dog001.jpg
dog002.jpg
...
cats/
cat001.jpg
cat002.jpg
...为了避免重复占用磁盘空间，这里采用链接文件夹的方法
In [6]:
import shutil
#训练集分为猫狗两个种类
train_filenames = os.listdir('train')
train_cat = filter(lambda x:x[:3] == 'cat', train_filenames)
train_dog = filter(lambda x:x[:3] == 'dog', train_filenames)
if isdir('train2'):
shutil.rmtree('train2')
os.makedirs('train2/cat')
os.makedirs('train2/dog')
abspath = os.getcwd()
for cat in train_cat:
os.symlink(os.path.join(abspath, 'train/' + cat), 'train2/cat/' + cat)
for dog in train_dog:
os.symlink(os.path.join(abspath, 'train/' + dog), 'train2/dog/' + dog)
#测试集
if isdir('test2'):
shutil.rmtree('test2')
os.makedirs('test2')
os.symlink(os.path.join(abspath, 'test'), 'test2/test')
print('symlink created.')
三、预训练模型输出类别¶
ImageNet(14,197,122 images, 21841 synsets indexed)是一个非常巨大的数据集，keras 中包含的预训练集就是在这个数据集上训练完成的，输出的结果中包含1000个分类（ https://s3.amazonaws.com/deep-learning-models/image-models/imagenet_class_index.json ）。 其中猫和狗的类别如下：
狗：151("n02085620", "Chihuahua") ~ 276("n02117135", "hyena")
猫：281("n02123045", "tabby") ~ 287("n02127052", "lynx")
In [7]:
dog_classes = [
'n02085620','n02085782','n02085936','n02086079'
,'n02086240','n02086646','n02086910','n02087046'
,'n02087394','n02088094','n02088238','n02088364'
,'n02088466','n02088632','n02089078','n02089867'
,'n02089973','n02090379','n02090622','n02090721'
,'n02091032','n02091134','n02091244','n02091467'
,'n02091635','n02091831','n02092002','n02092339'
,'n02093256','n02093428','n02093647','n02093754'
,'n02093859','n02093991','n02094114','n02094258'
,'n02094433','n02095314','n02095570','n02095889'
,'n02096051','n02096177','n02096294','n02096437'
,'n02096585','n02097047','n02097130','n02097209'
,'n02097298','n02097474','n02097658','n02098105'
,'n02098286','n02098413','n02099267','n02099429'
,'n02099601','n02099712','n02099849','n02100236'
,'n02100583','n02100735','n02100877','n02101006'
,'n02101388','n02101556','n02102040','n02102177'
,'n02102318','n02102480','n02102973','n02104029'
,'n02104365','n02105056','n02105162','n02105251'
,'n02105412','n02105505','n02105641','n02105855'
,'n02106030','n02106166','n02106382','n02106550'
,'n02106662','n02107142','n02107312','n02107574'
,'n02107683','n02107908','n02108000','n02108089'
,'n02108422','n02108551','n02108915','n02109047'
,'n02109525','n02109961','n02110063','n02110185'
,'n02110341','n02110627','n02110806','n02110958'
,'n02111129','n02111277','n02111500','n02111889'
,'n02112018','n02112137','n02112350','n02112706'
,'n02113023','n02113186','n02113624','n02113712'
,'n02113799','n02113978','n02114367','n02114548'
,'n02114712','n02114855','n02115641','n02115913'
,'n02116738','n02117135']
cat_classes = [
'n02123045','n02123159','n02123394','n02123597'
,'n02124075','n02125311','n02127052']
print("dog_classes: %s" % len(dog_classes))
print("cat_clasees: %s" % len(cat_classes))
四、利用预训练模型检出异常¶
In [8]:
from keras.models import *
from keras.preprocessing.image import *
from keras.applications import *
from tqdm import tqdm
def checkout_outlier(MODEL, preprocess=None, decode=None,
dirs='train2/cat', from_classes=cat_classes, size=(224,224), top=25):
model_name = MODEL.__name__
model = MODEL(weights='imagenet')
result = []
pbar = tqdm(os.listdir(dirs))
for file in pbar:
img = load_img('train/'+file, target_size=size)
x = img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess(x)
pred = model.predict(x)
decpred = decode(pred, top = top)
correct = False
for predicted in enumerate(decpred[0]):
if predicted[1][0] in from_classes:
correct = True
break
if not correct and file not in result:
result.append(file)
pbar.set_description('%s checking %s' % (model_name, file))
return result
cat 文件夹中被预测为不是猫的文件：
In [9]:
cat_outlier1 = checkout_outlier(ResNet50,
preprocess=resnet50.preprocess_input,
decode=resnet50.decode_predictions)
cat_outlier2 = checkout_outlier(InceptionV3,
preprocess=inception_v3.preprocess_input,
decode=inception_v3.decode_predictions,
size=(299,299))
#cat_outlier = [v for v in cat_outlier1 if v in cat_outlier2]
cat_outlier = list(set(cat_outlier1) & set(cat_outlier2))
print('cat outlier num: %d' % len(cat_outlier))
print(cat_outlier)
dog 文件夹中被预测为不是狗的文件：
In [10]:
dog_outlier1 = checkout_outlier(ResNet50, dirs='train2/dog', from_classes=dog_classes,
preprocess=resnet50.preprocess_input,
decode=resnet50.decode_predictions)
dog_outlier2 = checkout_outlier(InceptionV3,dirs='train2/dog', from_classes=dog_classes,
preprocess=inception_v3.preprocess_input,
decode=inception_v3.decode_predictions,
size=(299,299))
#dog_outlier = [v for v in dog_outlier1 if v in dog_outlier2]
dog_outlier = list(set(dog_outlier1) & set(dog_outlier2))
print('dog outlier num: %d' % len(dog_outlier))
print(dog_outlier)
五、猫狗异常值预览：¶
In [12]:
import random
fig = plt.figure(figsize=(20,15))
for i, name in enumerate(random.sample(cat_outlier, 21)):
a = fig.add_subplot(3, 7, i+1)
a.set_title(name)
a.axis('off')
img = load_img("train/"+name)
a.imshow(img, interpolation="nearest")
In [14]:
fig = plt.figure(figsize=(20,15))
for i, name in enumerate(random.sample(dog_outlier, 18)):
a = fig.add_subplot(3, 6, i+1)
a.set_title(name)
a.axis('off')
img = load_img("train/"+name)
a.imshow(img, interpolation="nearest")
In [15]:
#去掉异常值后的训练集，重新进行文件链接
if isdir('train3'):
shutil.rmtree('train3')
os.makedirs('train3/cat')
os.makedirs('train3/dog')
abspath = os.getcwd()
cat_plus = cat_outlier + tiny_pics
dog_plus = dog_outlier + tiny_pics
for cat in os.listdir('train2/cat'):
if cat not in cat_plus:
os.symlink(os.path.join(abspath, 'train/'+cat), 'train3/cat/'+cat)
for dog in os.listdir('train2/dog'):
if dog not in dog_plus:
os.symlink(os.path.join(abspath, 'train/'+dog), 'train3/dog/'+dog)
print('symlink created.')
print('train3/cat file nums: %d' % len(os.listdir('train3/cat')))
print('train3/dog file nums: %d' % len(os.listdir("train3/dog")))
In [16]:
np.save('tiny_pics.npy', tiny_pics)
np.save('cat_outlier.npy', cat_outlier)
np.save('dog_outlier.npy', dog_outlier)
print('saved.')
In [17]:
Image.open('train3/dog/dog.9220.jpg')
Out[17]:
